The Aegis of Abundance!

On work becoming optional, money becoming irrelevant, tax becoming

quaint, and the small matter of who pays for the interval

Future of Work series

Somewhere in a very large factory, in front of a very good interviewer, a prediction was

recently made.

Within five years, machine intelligence exceeds the sum of human intelligence. Within ten,

work becomes optional. Money stops mattering. Goods and services arrive in such volume

that the question of who can afford them dissolves. The state issues cheques not because

people are poor but because prices have stopped meaning anything.

The interviewer, to her credit, asked the obvious question. If money will not matter, why is

anyone buying your shares?

The answer was a shrug dressed as a syllogism.

Here is what should unsettle you.

The argument is not stupid.

That is the problem. A stupid argument can be ignored. A coherent argument with an

unexamined middle is far more dangerous, because it recruits your assent before it earns it.

And this one is coherent. Intelligence becomes cheap. Intelligence gets hands. Hands make

things. Things become plentiful. Money is a rationing device. Remove the need to ration and

you remove the need for the device.

Every step follows. Which is exactly why we should look very carefully at the joints.

Because what was offered that afternoon was not a forecast.

It was a prospectus.

And every prospectus has a risk factors section that nobody reads.

MOVEMENT ONE: THE PROMISEI. STATED AT FULL STRENGTH

Let me make the case properly, because an essay that only attacks a weak version of an

argument is not worth your Sunday.

The economy is the production of goods and the provision of services. Both are currently

bottlenecked by human cognitive and physical labour. Digital intelligence is dissolving the

first bottleneck at an astonishing clip. Embodied intelligence, in the form of general purpose

robotics, will dissolve the second. When both go, output is limited only by energy and

materials, and on any relevant timescale both are effectively unbounded.

Money exists to allocate scarce things. It is a rationing technology with excellent user

experience. In a world where the marginal cost of producing almost anything approaches

the cost of the electricity to produce it, rationing becomes a solution in search of a problem.

Prices collapse toward zero. Wages become irrelevant because purchasing power becomes

irrelevant. Tax becomes irrelevant because the state no longer needs to claim a share of a

constrained pot.

Work survives only as a hobby. You do not have to grow tomatoes. Some people grow

tomatoes.

Now the part almost nobody writes, including the sceptics who most need to.

II. THE UTOPIAN CASE, WHICH DESERVES A HEARING

Consider how strange the arrangement we currently defend actually is.

A human being arrives on this planet with a finite allocation of conscious hours. Somewhere

around eighty thousand of them, if things go well, are surrendered to an institution in

exchange for the right to eat. Not because the work is meaningful, though sometimes it is.

Because the alternative is destitution. We have built an entire moral vocabulary around this

arrangement, and the vocabulary is so complete that we struggle to see it as an

arrangement at all.

We call it earning a living. Read that phrase again as if you had never heard it.

Earning. A living.

Wage labour as the universal condition of adult existence is roughly two centuries old. It is

younger than the piano. Most humans who have ever lived did not have jobs; they had land,

obligations, seasons, kin, subsistence, and a great deal of misery, but not employment in

our sense. The idea that a person’s dignity is constituted by their position in a firm would

have been unintelligible to almost everyone who came before us.So when someone proposes that this arrangement might end, our reflexive horror is worth

examining. Some of it is prudence. Some of it is a failure of imagination so complete that it

has learned to call itself realism.

The strongest version of the promise is not that we get more stuff.

It is that the coercion goes out of the day.

That a person could spend their hours on what they actually judge worth doing, and that

this would not be a privilege of the wealthy but the ordinary condition of being alive. That

the enormous human talent currently absorbed by tasks nobody would choose could be

released into science, care, craft, argument, art, and the raising of children. That the phrase

“I have to go to work” would sound to our grandchildren the way “I have to go to the mill”

sounds to us.

If that were on offer, discomfort would be a poor reason to decline.

I want that on the record before I start pulling threads, because everything that follows is an

argument about the route, not about the destination.

III. AND THE DIRECTION IS NOT IN SERIOUS DISPUTE

One more concession, and it is a large one.

In July, Stanford’s Digital Economy Lab published an eighty-eight word statement. More

than two hundred economists and AI researchers signed it, sixteen of them Nobel laureates.

It warns that AI could drive an economic transformation larger than the Industrial Revolution,

arriving far faster, carrying both large-scale displacement risk and major gains in living

standards.

Among the signatories were Daron Acemoglu and Simon Johnson, who won the 2024 Nobel

partly for work arguing that technological shocks rarely produce the mass unemployment

people fear, and who have spent years publicly telling everyone to calm down.

Read that again.

The sceptics signed.

Note carefully what that is and is not. Acemoglu has said he has not abandoned his doubts

about the most aggressive industry timelines. This is not a conversion. It is a move from “the

claims are overstated” to “the uncertainty is now large enough to demand institutional

action”. That is a subtler thing than the headlines suggested, and a more useful one.

So the question is not whether something enormous is happening.

The question is whether the specific shape being promised, abundance without institutions,is the shape it takes.

It is not. And the reason is not moral. It is mechanical.

MOVEMENT TWO: THE MECHANISM

IV. THE CURVE THAT EXPLAINS WHY EVERYONE IS TALKING

PAST EACH OTHER

Start here, because this is the single most important thing in the entire debate and it is

almost absent from public discussion.

Anton Korinek and Donghyun Suh simulated the transition. What they found is that the path

is not a slope. It is a hump.

As automation advances, human labour first becomes more valuable, not less. Workers are

complements to an increasingly productive capital stock, and they are the scarce input.

Almost all the benefits accrue to labour. Wages rise. Employment holds. Every optimist is

vindicated by the data, repeatedly, for years.

Then, past a threshold, in their simulation somewhere around eighty per cent of tasks

automated, the abundance of capital and the sheer narrowness of what remains for humans

to do causes wages to plummet and the returns to shift wholesale to capital.

WAGES UNDER PROGRESSIVE AUTOMATION

==================================

W |

A | _______

G | ___/ \

E | __/ \

S | __/ \

| __/ \

| _/ \___________

+——————————————–

0% automation share 100%

^

|

the threshold, which nobody

can locate in advance and

everyone recognises afterSit with the shape of that for a moment, because it dissolves the entire optimist versus

pessimist argument.

They are not disagreeing about the world. They are standing at different points on the same

curve and describing what they see.

And here is the sentence that should stop you.

The good phase and the bad phase look identical from inside the good phase.

Rising wages, healthy employment, complaints about labour shortages, confident essays

about how every previous automation panic was wrong. That is not evidence against the

threshold. That is what the approach to the threshold looks like. The Korinek and Suh

simulation produces exactly this reassuring data right up until it does not.

Which means the reassurance is uninformative. Not wrong. Uninformative. It is precisely the

reading you would get either way.

This is also why “we have automated before and it was fine” is a weaker argument than it

sounds. Prior automation waves moved the boundary and stopped. The question is not

whether the boundary moves. It is whether it moves far enough, fast enough, to cross a

point that no previous wave approached.

Nobody knows. Anyone who tells you they know is selling something, in either direction.

But an economy that cannot see its own threshold and cannot know how close it stands is

an economy that should be building shock absorbers, not writing prospectuses.

The graph goes up. We cheer the graph. We frame it, print it, laugh the laugh of people

certain that a rise is proof of nothing but the skies. The graph goes up. That is the tell. It

goes up gently. Then it fell.

V. THE IDEA AT THE CENTRE: THE LEDGER GOES DARK

Now to the claim I think matters more than any other in this essay, and which I have not seen

made elsewhere.

The proposition that money stops mattering treats money as one thing. It is at least three.

Money is a medium of exchange. Money is a store of value. And money is a ledger of

claims: a public, portable, legible record of who is owed what by whom.Abundance might genuinely erode the first. It complicates the second. It does absolutely

nothing to the third, because abundance does not abolish claims. It abolishes only the units

in which claims were denominated.

Regular readers will recognise the Two Ledgers here, and this is the sharpest form of that

argument I have yet found.

If prices no longer allocate the things that remain scarce, something else must. And

whatever does the allocating becomes the most powerful institution in that society,

precisely because its decisions are no longer disciplined by a public number that anyone

can check.

Price is a terrible allocation mechanism. Regressive, noisy, often cruel.

It is also legible. You can see it. Contest it. Tax it. Index it. Point at it in a parliament.

Replace it with allocation by algorithm, by queue, by eligibility, by relationship, by favour,

and you have not removed the ledger. You have made it unauditable.

Abundance does not abolish the ledger. It turns out the lights in the room where the

ledger is kept.

This is the part of the prospectus waved through fastest and deserving the most scrutiny.

The promise is framed as liberation from the tyranny of the price tag. What it actually

proposes is the removal of the last universally readable signal in economic life, at exactly

the moment when the entities doing the allocating become too complex for any human to

inspect.

And yet.

There is a decent counter and I will give it. Much of what matters most is already allocated

without prices. Love. Citizenship. Organ transplants. Places at state schools. Emergency

care in most of the developed world. We have non-market allocation and some of it works

rather well, precisely because we decided certain goods should not be rationed by

willingness to pay.

Fair. But notice what makes those systems tolerable. Every one is governed by an institution

with a mandate, a published rule, an appeals process, and someone you can vote out.

Non-market allocation without institutions is not abundance.

It is patronage.

VI. THE WEAK LINK

Now the least philosophical and most stubborn objection.Charles Jones and Christopher Tonetti have a phrase for it: weak links tame the growth

explosion. Growth is constrained not by what you do brilliantly but by what is essential and

hard to improve. One bottleneck throttles the system, no matter how spectacular the rest of

it is.

The current weak link is electricity.

The International Energy Agency’s base case has global data centre consumption roughly

doubling to around 945 terawatt hours by 2030, close to three per cent of world demand,

reaching about 1,200 TWh by 2035, with a higher scenario past 1,700 TWh. Data centre

demand grew seventeen per cent in 2025; AI-focused facilities surged around fifty per cent.

Three per cent of global demand is not apocalyptic, and the IEA says so plainly. But

aggregate share is the wrong lens. These loads cluster, and they cluster where the grid is

already congested. The Agency notes that developers blocked by slow grid connections are

building onsite gas generation, and that meeting variable data centre load reliably that way

requires overbuilding generation by thirty to seventy per cent. One analyst house projects

power shortages constraining forty per cent of AI data centres by 2027.

So the quasi-infinite economy is currently queuing for a transformer.

Now the honest counter, because this is where I could be wrong.

Efficiency is improving fast. Cost per unit of capability has fallen by orders of magnitude.

Algorithmic progress has repeatedly outpaced hardware. A bottleneck today is not

necessarily a ceiling.

But efficiency cuts the other way too, and the mechanism has a name. Jevons observed in

the 1860s that improving the efficiency of coal use increased total coal consumption,

because cheaper meant more used. Every efficiency gain in inference has so far been met

with more inference, longer reasoning chains, more agents, more calls. Efficiency is not a

solution to the constraint. Historically it is an accelerant of it.

Either way, one fact holds.

Energy requires infrastructure. Infrastructure requires permits. Permits require politics.

The path to a world without institutions runs directly through the institutions.

You cannot build past the state on your way to a world where the state is unnecessary. The

state is standing in the substation.

VII. WHAT ABUNDANCE CANNOT REACH

Second weak link, and it is older than the grid.When one sector becomes dramatically more productive and another does not, the

unproductive sector does not become cheap. It becomes the whole economy. William

Baumol identified this in the 1960s and nobody has repealed it. Manufacturing productivity

rises, wages rise across the labour market because workers can move, and so the string

quartet costs more, not less, measured against everything else.

Apply it to a world where cognition is nearly free.

THE ABUNDANCE GRADIENT

======================

CRASHES TOWARD ZERO | RESISTS, THEN ABSORBS

| THE FREED SPENDING

—————————+—————————

Text and code | Land and location

Analysis and synthesis | Care, presence, touch

Design iterations | Trust and vouching

Translation | Legitimacy and consent

First drafts of everything | Time with specific people

Routine diagnosis | Scarce physical inputs

Tutoring at scale | Political permission

—————————+—————————

Price falls. Volume | Price rises to absorb the

explodes. | purchasing power released

| by the left-hand column

Philippe Aghion has shown that combining cost disease with automation can produce

balanced growth with a rising capital share. So this is not a knock-down objection to

abundance.

And here is the concession the sceptics owe.

Cost disease is fatal to the growth statistic, not necessarily to human welfare. If the

fast sectors become so abundant that their prices collapse, you can be enormously richer in

every way that matters while GDP growth looks flat and disappointing. A world of stagnant

measured growth and extraordinary material plenty is entirely possible, and it would look like

failure in every chart and like success in every life.

A related caveat on positional goods, which I have previously leaned on too hard. Yes, status

is scarce by construction; you cannot manufacture the better address or the seat at the

table. But positional competition exists at every level of wealth and has never previously

prevented enormous welfare gains. Someone competing for status in 1850 did so while

dying of cholera. Someone competing for status in 2050 may do so in perfect health with

unlimited leisure.Positional goods prove that inequality persists.

They do not prove that abundance fails.

That distinction matters, and I got it wrong before.

VIII. THE EVIDENCE SO FAR IS EMBARRASSING FOR EVERYONE

Task-level gains are real and large. Depending on study and task, measured improvements

run between fourteen and fifty-five per cent.

Organisational gains are, so far, close to invisible. An MIT research group reported around

ninety-five per cent of enterprise generative AI pilots never reach production. A National

Bureau of Economic Research survey of roughly six thousand chief executives and finance

chiefs across four countries found the great majority reporting little operational impact, with

about ninety per cent seeing no measurable productivity improvement, and perceived gains

consistently exceeding measured gains. PwC’s 2026 Global CEO Survey, covering more

than four thousand chief executives across ninety-five countries, found a majority saying

they had got nothing from their AI investments.

Meanwhile the capital expenditure is historic. Hyperscaler AI-related spend moved from

roughly $235 billion in 2024 to around $400 billion in 2025, with 2026 estimates past $700

billion. In the first quarter of 2026, AI-related investment accounted for something in the

region of three quarters of American GDP growth.

Now, two honest caveats, because I have been sloppy about both.

First, that GDP growth figure describes a contribution at a moment of otherwise weak

demand. It is not by itself evidence that the investment is unproductive. Building a railway

shows up as construction before it shows up as freight.

Second, this is precisely what the last general purpose technology looked like at this stage.

The productivity paradox named in 1987 took roughly eight more years to resolve, and when

it resolved it resolved spectacularly. Erik Brynjolfsson calls this the productivity J-curve:

intangible reorganisation costs, retraining, process redesign, depress measured productivity

before they lift it. Firms are investing in things the accounts cannot see.

The J-curve is not a rebuttal of my position. It is the mechanism behind it. The Adoption

Redesign Gap is exactly the dip in the J.

But notice what the concession does to the timeline.

If the gains arrive in 2032, the abundance arrives later still, and the labour market disruption

arrives first. The costs are front-loaded and concentrated. The benefits are back-loaded

and diffuse.Every transition is a transfer before it is a dividend.

IX. THE ECONOMISTS’ BENCH

Time to bring in the people who study this for a living, because the public conversation

about income floors is roughly a decade behind the professional one.

Daron Acemoglu and Simon Johnson offer the most forceful opposition to universal basic

income, and it is not a fiscal objection. Acemoglu argued in 2019 that UBI is not merely

unrealistic and ineffective but suspect on democratic grounds, and in Power and Progress

he and Johnson sharpen this into the charge that it is defeatist. It concedes that economic

contribution will come from a shrinking fraction of society and that everyone else must be

paid off. His comparison is to bread and circuses. His alternative: redirect technological

change away from automation toward worker complementarity, rebuild countervailing power

through worker organisation, and use conventional instruments, universal healthcare, better

unemployment insurance, serious retraining, an expanded earned income tax credit, which

he notes already functions as a guaranteed income for low-wage workers at far lower cost

while rewarding work.

His critics fairly observe that the book disposes of UBI in about two pages and engages

almost none of the literature on conditionality, poverty traps and take-up.

Anton Korinek is the most operationally useful voice in the field. His position: we do not

need a UBI now and should not want one now, because too many tasks still require humans.

But if transformative AI arrives quickly, we will want a distribution system already built.

Hence his seed UBI: small payments now that establish the identity, payments and political

infrastructure, scalable later if and when the funding base materialises. With Lee Lockwood

he has written a primer on public finance in the age of AI, examining what to tax when AGI

erodes the labour income and consumption bases, with a candidate set spanning tokens,

robots, robot services and digital services.

Build the pipe before you need the water. That is the single most sensible sentence available

on this subject.

David Autor supplies the counterweight I owe you. His distinction is between two kinds of

tool. An automation tool eliminates expertise. A collaboration tool is a force multiplier

for expertise. Computerisation, he argues, was a gift to elite experts and a wrecking ball to

mass expertise, because information turned out to be merely an input to the consequential

function, which is decision-making, and decision-making belonged to elites. AI’s unique

opportunity is to push expertise back downward: to let workers with complementary

knowledge perform higher-stakes decisions currently reserved for doctors, lawyers, senior

engineers and professors.He is explicit that this is not a forecast but an argument about what is possible, and that the

outcome is not deterministic. How AI is designed is a choice. The institutions deciding who

benefits are guilds, licensing bodies, firms and education systems, and he notes that

restrictive professional institutions currently protect elite incumbents.

Yanis Varoufakis, from the opposite end of the spectrum, argues labour markets are

beyond reform and that no wealth or income tax touches the problem. His alternative is a

Universal Basic Dividend funded from a public equity stake rather than taxation. A trust

fund for the many. Not paying people to be idle, in his framing, but releasing them from the

necessity of selling their labour.

The sceptical flank matters too. A recent American Enterprise Institute review of 122 basic

income pilots found only 52 with published outcomes, 35 randomised, 30 reporting

employment. Across those 30 the mean effect was plus 0.8 percentage points on

employment. Among the four largest, over half of all participants, it was minus 3.2. Their

conclusion: the evidence base is thin, badly timed, and a poor foundation for a permanent

entitlement. The Tax Foundation adds that if AI genuinely raises productivity, revenues rise

with it, and the claimed automation bias in the tax code is contested, since firms can deduct

wages fully and immediately while capital deductions have historically been spread over

years.

Now the paragraph nobody is writing.

Almost no serious economist currently supports an unconditional universal basic income as

a first-best policy for today’s economy. The opposition comes from the left, which calls it

defeatist, from the right, which finds the evidence weak and the incentives wrong, and from

the technocratic centre, which finds it badly targeted relative to alternatives.

Support instead clusters around three things: contingent mechanisms built now and

triggered later, capital stakes and dividends on an ownership logic, and wage insurance plus

services that target the actual harm at the actual moment.

The serious debate has already moved past UBI. The public one has not noticed.

X. JUDGEMENT: CONCENTRATED, OR DISTRIBUTED?

Which brings us to the tension I want to stage rather than resolve, because it is the real

question and I do not think it is settled.

My position, argued across this series, is that judgement is conserved. Automating a task

does not destroy the judgement that task required; it relocates it. Someone must still decide

whether the output is right, whether the question was the right question, whether the edge

case is an edge case or the start of a catastrophe. As machine output rises, the judgement

required rises with it.So work does not become optional.

Work becomes concentrated.

Into fewer people, holding more of it, at higher stakes, with less of the apprenticeship ladder

that used to produce people capable of holding it. This is Seniorisation, and it is now visible

rather than merely arguable. Stanford’s work on employment records shows software

developers aged twenty-two to twenty-five down close to twenty per cent from a 2024

peak, while developers aged thirty and above at the same firms grew between six and

twelve per cent.

The ladder is not collapsing from the top. It is being sawn off at the bottom.

Autor’s thesis points the other way. If AI is built as a collaboration tool, judgement gets

distributed, not concentrated. The nurse practitioner does more of what the physician did.

The paralegal does more of what the partner did. The middle rebuilds.

Both cannot be true in the same sector at the same time.

And the difference between them is not technological. It is a matter of design and

institutional choice, which is exactly Autor’s point and exactly why he refuses to call it a

forecast.

Which means Seniorisation is not a prophecy.

It is what happens by default when nobody chooses otherwise.

That is a considerably more useful claim than the fatalistic version, and I am revising toward

it.

XI. ONE PERSON, SO THAT THIS IS ABOUT SOMEBODY

Everything above is aggregate. Let me put one person in the room, composite but entirely

ordinary.

Wei Ling is twenty-four. Second year associate. Two years ago the first year of her job was

reconciliations, tie-outs, sampling, chasing schedules, and the slow accumulation of a feel

for when a number is lying. Tedious, largely automatable, and quietly the mechanism by

which she was going to become senior.

Most of that work now runs without her.

Her firm is delighted. Her chargeable hours look better than any second year in the firm’s

history. She is doing work that used to belong to someone three years ahead of her, and she

is doing it competently, which everyone takes as evidence that the transition is going well.Here is what nobody is measuring.

She is making judgements she has not yet earned the right to make, and she knows it. The

reviewer above her is stretched across more juniors than before. The feedback loop that

used to catch her is thinner. When she is wrong, she will be wrong at a level where being

wrong matters, without the three years of being wrong cheaply that used to precede it.

And the person who would have been hired below her, the one who would have done the

reconciliations and started the same climb, was never hired at all. No redundancy. No

headline. No line in any labour force survey.

That is a Shadow Job, and Wei Ling is standing on the space where it should have been.

Multiply her by a cohort and you have the entire argument of this series in one office.

XII. LABOUR, WORK, AND THE THING NOBODY IS

AUTOMATING

Now the deeper question, because the employment statistics are a proxy for something

they cannot measure.

Hannah Arendt drew a distinction we badly need here. She separated labour, the cyclical

activity of sustaining life, which produces nothing that lasts and must be done again

tomorrow. Work, the fabrication of durable things, which leaves an object in the world that

outlives the making of it. And action, the political and speech-based activity through which

we appear to one another, disclose who we are, and do things that could not have been

predicted.

Automation is coming for the first two.

Labour goes first and largely deserves to. Nobody should mourn the reconciliations.

Work is next, and its loss is more serious than we admit, because fabrication confers a

relation to the object. The gardener’s tomatoes are still hers. This is where the hobby

analogy quietly fails: the question is not whether people will have activities. It is whether

their activities will produce anything that stands independently of them, or whether we are

all to be permanently entertained consumers of things we did not make.

But action is untouched. Nothing in this technology automates appearing before others,

arguing, persuading, deciding together, taking responsibility for a decision that could have

gone otherwise.

And here is the thing that should worry you most.

A society can be materially abundant and politically impoverished at the same time. Infact that combination is easier to produce than any other, because abundance removes the

friction that used to force people into public rooms with each other.

If the abundant future frees us from labour and work while hollowing out action, the

employment statistics will look fine and something far more important will have been lost.

The essay’s central worry, restated in the vocabulary it needed: the danger is not that we

run out of jobs.

It is that we run out of reasons to show up.

MOVEMENT THREE: THE FORK

XIII. THE FLOOR: WHAT THE EVIDENCE ACTUALLY SHOWS

Both sides cite the trials. Neither side reads them.

The largest American trial. Three years, 1,000 low-income participants receiving $1,000

monthly against 2,000 controls receiving $50. Participation fell about two percentage

points. Hours fell roughly 1.3 to 1.4 per week, with partners reducing by a comparable

amount. Total individual income excluding the transfer fell around $1,500 a year. Spending

rose about $310 monthly, concentrated in food, rent and transport, with increased support

to friends and family. Recipients were more likely to seek medical care, to move, to search

for a different job, to attempt entrepreneurship. No significant improvement in physical

health. No measured improvement in job quality. Attitudes toward the value of work did not

decline. Researchers characterised the labour response as increased agency rather than

withdrawal.

The wider literature. The AEI figures above: plus 0.8 percentage points on average across

thirty randomised pilots, minus 3.2 across the four largest.

That minus 3.2 needs its caveat, and here it is. The large trials differ from the small ones in

more than size. Timing, population, transfer size and pandemic-era labour conditions all

confound. It is a serious number and it is not a clean one.

The one permanent universal programme. Alaska has paid an annual dividend from a

sovereign fund since 1982. Synthetic control analysis found no effect on aggregate

employment and a 1.8 percentage point increase in part-time work, consistent with cash

stimulating local demand. Poverty effects have been meaningful, particularly for indigenous

and elderly Alaskans. Payments typically $1,000 to $2,000, peaking above $3,200 in 2022,set at $1,200 for 2026.

Now the sentence that matters more than any number above.

None of these experiments test the proposition being made.

Every trial gives money to some people inside an economy where jobs exist, prices are set

by markets, and the recipients’ neighbours are still working. That is an island experiment. It

tells you what a floor does to an individual. It tells you almost nothing about what a floor

does to a society whose labour market has been hollowed out.

The general equilibrium effects, on prices, rents, norms, coalitions, and the meaning people

attach to their days, are exactly what a randomised trial cannot measure.

You cannot pilot the end of work.

Which is precisely why Korinek’s seed proposal is the right instinct. If you cannot test the

mechanism, build it small, keep it live, and be ready to scale it.

XIV. THE FLOOR HAS VARIANTS AND THEY ARE NOT

INTERCHANGEABLE

INSTRUMENTS FOR A WORLD WITH LESS PAID WORK

===========================================

INSTRUMENT WHAT IT IS OPTIMISES FOR TRADES AWAY

—————— —————— —————- —————

Universal Basic Flat unconditional Simplicity, no Cost. Pays the

Income cash to all stigma, no rich too. Weak

eligibility trap targeting.

Seed UBI Small universal Having the Politically odd

payment built now, MECHANISM ready at small size.

scalable later before the need Easy to cancel.

Negative Income Cash tapering as Targeting, lower Marginal rate

Tax income rises headline cost cliffs. Means

testing returns.

Guaranteed Income Cash to a defined Political Stigma. Divides

group, often time feasibility, deserving from

limited cost control undeserving.

Universal Basic Free or near-free Insulating the Huge state

Services care, health, RIGHT-HAND capacity needed.transport, housing COLUMN from Quality risk.

price inflation Less freedom.

Job Guarantee State as employer Structure, Make-work risk.

of last resort dignity, skill State must invent

formation useful work.

Citizens Dividend Payout from a Tying income to Fund must exist.

publicly owned OWNERSHIP, not Returns vary.

asset stake charity Slow to build.

Universal Basic Capital endowment A stake, not a Sequencing.

Capital at adulthood stream Can be lost.

Wage Insurance Time-limited The actual harm Nothing if

displacement pay at the actual displacement is

plus retraining moment permanent

One row deserves separate attention, and the difference is not economic. It is moral.

A transfer has a direction. Someone gives you money. Direction creates standing. You are,

structurally, a recipient.

A dividend has no direction. It is a return on an asset you own. You are, structurally, an

owner.

Alaska’s fund has survived four decades in one of the most fiscally conservative

jurisdictions in the developed world. Not because Alaskans are unusually communitarian,

but because the payment is understood as theirs. Attempts to cut it are political suicide.

Compare that durability with almost any welfare programme anywhere.

Varoufakis arrives at the same architecture from the far left. The Alaskan legislature arrived

at it from the right. When two ends of the spectrum converge on a design, the design is

usually telling you something.

If you want an income floor that survives a change of government, do not build a transfer.

Build an ownership stake.

XV. THE ARITHMETIC, WHICH IS UNKIND

To lift every American below the federal poverty threshold up to it requires roughly $573

billion a year. Total dividends paid by the entire S&P 500 in 2025 were around $665 billion.

That is every dividend, from every listed company, in every sector, to cover one country’s

poverty gap alone.The AI firms themselves, semiconductors aside, are largely not yet profitable. They are

spending on a scale that dwarfs their revenues, financed on expectations that have not yet

appeared in any national statistic.

So the proposal, stated plainly: fund universal high income from the profits of firms that do

not yet make profits, on infrastructure not yet paid for, generating productivity that does not

yet appear in the accounts.

A plan requiring a miracle in the middle is not a plan. It is a wish with a spreadsheet

attached.

And the Alaskan precedent contains a detail its enthusiasts skip. Dividends have averaged

under $2,000 annually and have trended down. A well-run sovereign fund built on a

genuinely world-class resource endowment, over four decades, delivers a payment

meaningful for a poor household and nowhere near sufficient to make work optional for

anyone.

The dividend is real. It is just small.

Plan accordingly.

XVI. WHAT TO TAX WHEN THE BASE DISSOLVES

Developed economies tax labour. Income tax and payroll contributions are the load-bearing

wall of almost every treasury on earth. If the labour share falls, the base erodes precisely

when demand for state support rises. The squeeze arrives from both ends at once.

The candidate instruments, briefly.

Automation and robot taxes. Levy the machine to restore neutrality with labour. Intuitively

appealing, administratively brutal. What counts as a robot? Was the ATM? South Korea’s

celebrated “robot tax” was actually a trimming of automation investment credits, which tells

you how hard the direct version is.

Token taxes. Levy the billed token, already metered and invoiced. Elegantly auditable. But it

penalises verbose architectures over efficient ones, taxing chattiness rather than value, and

it is arbitraged by moving inference on-premises or offshore.

Compute or FLOP taxes. Levy the substrate. Highly auditable, since large training runs are

visible from the electricity meter. But it taxes the frontier most heavily precisely where you

want responsible actors rather than jurisdictions that will not govern at all.

Land. Here is the instrument the debate keeps skipping, and it follows directly from the

abundance gradient. If cheap cognition releases enormous purchasing power that then

flows into immobile, non-reproducible assets, the surplus ends up capitalised into landvalues. That is the textbook definition of an economic rent. Taxing it is the one instrument

that is simultaneously distortion-minimising, impossible to offshore, and precisely aimed at

where the abundance actually lands.

It is also politically radioactive, which is why nobody proposes it and everybody should.

Broad base reform. Reduce the payroll wedge, remove preferential treatment of

automation capital, shift toward capital, rents and land. Dull. Robust. Does not require

anyone to define artificial intelligence in primary legislation, which will otherwise consume a

decade in litigation.

Now the contradiction at the heart of the prospectus.

It says the state becomes unnecessary because tax becomes unnecessary. Yet it proposes

the Treasury issues cheques to the entire population, indexed to output, forever.

Consider what that requires. Identity coverage for every resident. Payments infrastructure of

extraordinary reliability. A body setting the level, the indexation, the residency rules, the

treatment of children, migrants, prisoners, the recently deceased. A mechanism for

capturing the surplus. Enforcement. A court for appeals. A legislature to change it when it

goes wrong.

A state that distributes abundance is a larger state than one that merely taxes

scarcity.

You do not get to abolish the institution and simultaneously assign it the largest logistical

operation in human history.

XVII. FOUR LEVERS, NOT ONE

The interview’s most interesting proposal was about safety, and it has been almost entirely

overlooked. The leading labs hold regular calls. They give each other one to two weeks of

early API access before release so competitors can probe for danger. Government becomes

an escalation path, invoked only when a lab refuses to address a demonstrated risk.

Let me steelman it, then place it properly.

Lever one: self-regulation. The case is strong. The technology iterates in weeks,

legislation in years. The people capable of finding a novel failure mode largely work at

frontier labs. The lab knows what its model can do; the regulator knows what the lab said. In

1975 molecular biologists imposed a voluntary moratorium on recombinant DNA, convened,

agreed containment, resumed, and it worked. Adversarial evaluation by a party with skills

and an incentive to embarrass you is more rigorous than self-assessment.

And credit where it is owed: frontier safety frameworks, capability thresholds and if-thencommitments represent more safety infrastructure than any law required, and more than

comparable industries built voluntarily at similar stages.

The limits are structural rather than moral. The second International AI Safety Report found

twelve companies had published or updated frontier safety frameworks, most measures

voluntary, effectiveness uncertain given limited external review and uneven compliance. One

cited study found companies had honoured previous voluntary commitments inconsistently:

highest adherence on content provenance and watermarking, lowest on securing model

weights.

Read that ordering again. Compliance strongest where cheapest and most visible. Weakest

where most expensive and least visible.

That is not a claim about anyone’s integrity. It is what voluntary regimes do under

competitive pressure, in every industry, every time.

Two further objections. The film ratings board so often cited as the model was not created

by conscience; it was created under credible, imminent threat of statutory censorship. Self-

regulation works in the shadow of the statute, and remove the shadow and you have a trade

association. And ratings governed the distribution of finished artefacts. A frontier model is

not an artefact but a general capability, an unbounded space of uses nobody has

enumerated. Two weeks of competitor access probes the fraction competitors think to look

for, which correlates suspiciously well with the fraction they are themselves exposed to.

Lever two: statute. Necessary, slow, and only as good as the capacity behind it, which

brings us to the last fortnight.

Lever three: liability. The most powerful lever in the entire system, and the one nobody

discusses.

Regulation moves at the speed of parliaments. Liability moves at the speed of the first

successful claim, and it reaches conduct no statute anticipated. Product liability.

Professional negligence. Duty of care. Vicarious responsibility for the actions of an

autonomous agent deployed on your behalf. And behind all of it, the insurance market,

which prices risk continuously, demands evidence, and withdraws cover from behaviour it

cannot underwrite.

If you want a mechanism that disciplines frontier behaviour without requiring a regulator to

out-hire a laboratory, this is it. Insurers do not need to understand transformer

architectures. They need loss data and the right to refuse.

There is also a live legal question underneath the whole ownership argument. Whether

training on copyrighted and personal data requires licence, compensation, or neither, is

being litigated right now across multiple jurisdictions. That litigation determines whether

“the training corpus was unpriced labour” is a moral claim with a courtroom or a moral claimwithout one.

Lever four: international red lines. Verifiable prohibitions on the uses nobody should

permit. Hardest to achieve, easiest to defect from, and the only lever that addresses risks

which do not respect borders.

Four levers. Every serious governance conversation uses one and forgets the other three.

XVIII. THE FORTNIGHT THAT PROVED THE POINT

On 27 July 2026, days ago, the European Union’s Digital Omnibus entered into force.

Precision matters here, so let me be exact. Obligations for general-purpose AI models

commenced in August 2025 and were not deferred. Article 50 transparency obligations

largely applied on the original date, with a short grandfathering window for marking legacy

systems. New prohibitions were added, including on non-consensual intimate imagery and

child sexual abuse material. The AI Office gained expanded supervisory powers over

general-purpose models.

What moved was the high-risk regime. Standalone Annex III systems now face 2 December

2027 rather than 2 August 2026. AI embedded in regulated products under Annex I moves

to August 2028. National regulatory sandboxes slip a year.

And Annex III covers employment. Education. Access to essential services. Credit scoring.

The rules governing AI’s effect on work were the ones postponed.

The reason was not malice, and the honest version is more damning than the cynical one.

Harmonised standards were not finalised. National competent authorities were not

designated. Compliance tooling did not exist. The regulation was written faster than the

capacity to implement it could be built.

That is not an argument against regulation. It is an argument about sequencing, and it is

the most useful lesson available to any government reading this. Legislation without

implementation capacity produces a deadline, then a delay, then a credibility loss. The delay

is not the failure. It is the visible symptom of a failure two years earlier, when someone

drafted a rule without funding the body that would enforce it.

Regular readers know I treat trust as an operating system rather than a sentiment. This is

what a kernel panic looks like in the log. The rules did not fail because they were wrong.

They failed because the runtime was never installed.

XIX. THE ECONOMY THAT NEVER TOUCHES A HUMANOne more development, because it extends the ledger argument to its conclusion.

Agents now discover, negotiate, purchase and pay without a human in the loop.

Programmatic advertising and algorithmic trading were the precursors. 2026 is the year it

generalised into procurement, underwriting, compliance and customer service.

Measurement is being invented in real time. One analyst house estimated roughly thirty-six

billion dollars of annualised “agent GDP” as of July 2026, defined as net economic value

added by deployed agents, adjusted for overlap with human work and agent cost of goods

sold. The World Economic Forum has argued the agent economy requires a Know Your

Agent framework alongside Know Your Customer, warning that one in four enterprise

breaches by 2028 could stem from agent exploitation.

Consider what that does to Ghost GDP.

If agents transact with agents, the volume of economic activity that never touches a

household, a payroll, or an income tax return grows without limit. The national accounts

were built for an economy in which value passes through people. That assumption is being

quietly retired, and nobody has announced it.

And it converts TrustOS from a metaphor into an engineering requirement. In an economy of

anonymous autonomous counterparties, trust stops being a cultural asset and becomes a

protocol.

XX. WHAT A RULE-TAKER DOES

Now the question that matters most to readers in this region, and which almost no one is

writing about.

Most economies in Asia Pacific are neither rule-makers nor frontier builders. The standards

will be written in Brussels. The frontier will be built in a handful of American and Chinese

laboratories. The compute will be sited where the power is. For a small, open, trade-

dependent economy, the entire preceding essay is a description of weather, not of policy.

So what does a rule-taker actually do?

Arbitrage the standards rather than duplicating them. Do not write your own AI Act.

Recognise the strictest applicable regime, build conformity infrastructure for it, and sell that

conformity as a service to firms that need to export into it. Being the jurisdiction where

compliance is cheapest and most credible is a real industrial policy.

Compete on implementation capacity, not on legislation. The Brussels lesson cuts both

ways. If the binding constraint globally is the shortage of bodies able to assess, audit and

certify, then the country that builds that capacity first exports it. Regulatory capability is atradeable good and nobody is currently treating it as one.

Own a stake in the physical layer. Compute follows power, land and water. A jurisdiction

that controls grid access, cooling and planning consent has genuine negotiating leverage,

and the leverage is highest now, during the buildout, not later when the returns arrive.

Equity, offtake rights, sovereign compute allocations. Ask for something more durable than

jobs.

Build the transition apparatus before the transition. Singapore already has more of this

than almost anywhere: SkillsFuture credits, mid-career top-ups, the Jobseeker Support

scheme paying up to $6,000 over six months, front-loaded and conditional on active search,

with SkillsFuture and Workforce Singapore merging into a single agency this year and an AI

readiness self-diagnostic added to the national portal. Over six hundred thousand people

took supported training in 2025.

None of this is glamorous. All of it is the pipe being built before the water.

And measure what nobody else measures. A small state can lead the world in statistics.

Entry-level hiring rates by sector and age. Time-to-competence within firms. Roles not

created. Whoever produces the first credible Shadow Jobs series will define the terms of

the global debate for a decade, and it costs a rounding error.

The rule-taker’s advantage is not scale.

It is that a small state can decide something on Tuesday and have it running by Friday.

XXI. THE OBJECTION I CANNOT FULLY ANSWER

Here is the strongest argument against everything I have written, and it comes from history

rather than ideology.

Factory acts. Trade unions. Social insurance. Public education. The welfare state. Not one

was built in anticipation. Every single one was built during or after the crisis it addressed, by

people radicalised by the experience of it. The eight hour day was not designed by

farsighted planners in 1780. It was won by people who had already spent their childhoods in

mills.

Institutions are, historically, crisis products.

If that pattern holds, then “act now” is a well-intentioned category error, and the honest

forecast is that we will build the institutions after the damage, exactly as we always have,

and the essays urging foresight will be quaint artefacts of the period before anybody was

hurt.

I cannot dismiss this. It is the historical record.Here is the only answer I have, and you should weigh it yourself.

Every previous transition ran over two to three generations. Time was the mechanism that

made reactive institution-building work: the crisis arrived slowly enough that the response

could mature while the damage accumulated. The factory system had eighty years to

produce the factory acts.

The claim on the table is that this one arrives in under one working life.

If that claim is even half right, the reactive model fails not because it is unwise but because

the clock does not permit it. There is no eighty years. There is possibly no eight.

Which means the historical pattern is not a reassurance.

It is the thing we have to break.

And notice what that requires. Not better forecasting. Institutions that can act under

genuine uncertainty, before the evidence is unambiguous, knowing they may be acting too

early. That is a capability almost no democracy currently possesses, and building it is a

harder problem than any of the policies in the next section.

XXII. WHAT TO EXPECT

My base case, so you can hold me to it.

Deflation on the left of the gradient, inflation on the right. Cognitive output collapses in

price. Housing, care, education, health and access to scarce human attention rise,

absorbing the freed purchasing power. Headline inflation looks confusing. Lived cost of

living worsens for renters and improves for owners. That gap becomes the defining political

cleavage of the 2030s.

Employment holds in aggregate; composition shifts violently. The unemployment rate is

a terrible instrument for detecting this and will mislead everyone for years. Watch entry-

level hiring, time-to-first-promotion, and the age profile within firms.

The fiscal squeeze arrives before the fiscal windfall. The most predictable crisis of the

next five years, and almost nobody is provisioning for it.

Political volatility becomes the binding constraint, not technology. Long before we learn

whether abundance is achievable, we learn whether the societies attempting it remain

governable during the attempt.

Institutional trust determines the outcome more than model capability does. A high-

trust society can redistribute, retrain and rebuild. A low-trust one cannot pass the legislation

to do any of it, and spends the decade litigating instead.XXIII. WHAT CAN ACTUALLY BE DONE

Diagnosis without prescription is entertainment. So: three horizons, with owners, and with

the opposition named, because a proposal that does not name its enemy is a document

destined for a drawer.

HORIZON ONE: THIS BUDGET CYCLE (finance ministry, no new law)

============================================================

ACTION OWNER WHO FIGHTS IT

—————————- ————— ——————-

Commission the Shadow Jobs Statistics Nobody. Do it now.

series: entry-level hiring office Costs a rounding

by age and sector, time-to- error.

competence, roles not made

Seed mechanism: small Treasury + Fiscal hawks, who

universal payment on revenue call it the thin

existing rails, live and authority end of the wedge.

scalable. Build the pipe. They are not wrong.

Fund enforcement capacity Regulator Treasury. Always

ahead of any AI statute. + finance the treasury.

Auditors, assessors, evals ministry

Wage insurance: generous, Labour Employers facing

fast, portable, time-limited ministry and conditional on search

levy. Manageable.

HORIZON TWO: THIS PARLIAMENT (primary legislation)

==================================================

ACTION OWNER WHO FIGHTS IT

—————————- ————— ——————-

Tax neutrality: cut the Treasury Capital-intensive

payroll wedge, end + revenue incumbents, and

preferential treatment of every lobbyist

automation capital they employ

Land value taxation of the Treasury Landowners. This

rents where abundance + local govt is the hardest

actually lands fight in politics

and the best idea

Sovereign stake in the Sovereign fund Shareholders,

physical layer: equity or + energy and trade lawyersofftake for grid access, ministry arguing state aid,

land and planning consent MFN and takings

Liability regime for Justice Deployers and

autonomous agents: ministry their insurers,

duty of care, vicarious until the insurers

responsibility, disclosure realise they win

Mandatory third-party Regulator Labs, on grounds

compliance review of of trade secrecy.

frontier safety frameworks Negotiable.

HORIZON THREE: TREATY AND COALITION (multi-year)

================================================

ACTION OWNER WHO FIGHTS IT

—————————- ————— ——————-

Verifiable international Foreign Everyone, until

red lines on the uses ministry someone defects

nobody should permit and proves the

point expensively

Mutual recognition of Trade ministry Nobody much. This

conformity assessment, is the rule-taker

exported as a service arbitrage.

Standards for measuring Statistical Inertia, which is

agent-to-agent activity coalition + undefeated across

in the national accounts central banks all of history

For institutions and firms. Stop measuring adoption; start measuring redesign. The gains

are real at task level and evaporate at organisational level, and that evaporation is a design

failure, not a technology failure. Treat apprenticeship as capital expenditure rather than

overhead: if you remove the junior tasks that produced senior judgement, you have deleted

your own succession pipeline and the bill arrives in seven years. Publish your seniorisation

map. And choose Autor’s fork deliberately: are you buying automation tools that eliminate

expertise, or collaboration tools that extend it? That is a procurement decision with a thirty

year consequence, currently being made by people optimising for this quarter.

Do it with them, not to them. HX equals CX plus EX is not a slogan. It is the observation that

you cannot deliver a good customer experience through an employee experience you have

hollowed out.

For individuals. Acquire judgement, not tasks. Own something, because holding only

labour in a world where returns shift to capital is a concentrated position in a decliningasset. And stay inside the institutions. The temptation during a legitimacy crisis is exit, and

exit is what produces the second branch of the fork. The people who show up to the

planning meeting, the union, the professional body, the select committee, are making the

better branch marginally more likely in the least glamorous way imaginable.

XXIV. THE FORK

Let me end where I always do, with one concession.

The optimistic vision is usually illustrated with a particular fictional future. A post-scarcity

federation. No money. Replicators. People working because they want to.

I am about to use it as an intuition pump rather than an argument, and I would rather say so

than pretend otherwise.

Everyone who cites it remembers the replicators.

Nobody remembers the rest.

That civilisation has an academy with a competitive entrance examination. A constitution.

Courts. A diplomatic corps. A chain of command. A prime directive its officers argue about

constantly and occasionally break at enormous personal cost. Rank. Discipline. Institutional

memory. A legal framework for artificial minds.

The replicator is not why that society works.

The institutions are why the replicator did not destroy it.

And note what those people spend their time doing. Not consuming. Not labouring. Not

even, mostly, fabricating. They spend it appearing before one another, arguing, deciding,

and taking responsibility for decisions that could have gone otherwise.

They spend it on action.

The other branch is not a technological failure either. The scarcity dystopia is not short of

machines. It is full of machines. It is short of law. What collapsed was not supply but the

capacity to make collective decisions about supply, at which point the machines carried on

working and simply changed whose interests they served.

Both futures have the technology.

Only one has the institutions.

That is the Fork. It has never been a fork between abundance and scarcity. It is a fork

between two ways of holding abundance, and the variable that decides it is not compute.Abundance is not the absence of institutions. It is the product of them.

Anyone selling you the first while dismantling the second is selling you the second future

with the first one’s brochure.

Every law that shields us now was carved by hands that shook, by people who had starved.

The mill came first. The act came forty years behind. That was the bargain time allowed

mankind. But this one comes inside a working life, too fast to learn its lesson from the knife.

So we must do the thing we have not done: build the institution before the harm has come.

The prospectus is on the table. The returns are speculative, the timeline aggressive, the risk

factors in small print, and the underwriters are also the auditors.

You are not being asked to believe it.

You are being asked to fund it, with a decade of your working life, on the promise of a

dividend nobody has yet demonstrated can be paid.

Read the small print.

Then go to the meeting.

Part of an ongoing series on the future of work. Sources referenced include the International

Energy Agency, the second International AI Safety Report, the European Union’s Digital

Omnibus and AI Act, Stanford’s Digital Economy Lab, the National Bureau of Economic

Research, MIT, the American Enterprise Institute, PwC’s 2026 Global CEO Survey, the World

Economic Forum, the Alaska Permanent Fund literature, Singapore’s SkillsFuture and

Workforce Singapore instruments, and published findings from the largest randomised

unconditional cash transfer studies to date. Work cited or engaged includes that of Daron

Acemoglu, Simon Johnson, David Autor, Erik Brynjolfsson, Anton Korinek, Donghyun Suh,

Lee Lockwood, Charles Jones, Christopher Tonetti, Philippe Aghion, William Baumol, Yanis

Varoufakis and Hannah Arendt. Wei Ling is a composite.